# Copied and adapted from: https://github.com/hao-ai-lab/FastVideo

# SPDX-License-Identifier: Apache-2.0
"""
Wan video diffusion pipeline implementation.

This module contains an implementation of the Wan video diffusion pipeline
using the modular pipeline architecture.
"""

from sglang.multimodal_gen.runtime.models.schedulers.scheduling_flow_unipc_multistep import (
    FlowUniPCMultistepScheduler,
)
from sglang.multimodal_gen.runtime.pipelines_core.composed_pipeline_base import (
    ComposedPipelineBase,
)
from sglang.multimodal_gen.runtime.pipelines_core.lora_pipeline import LoRAPipeline
from sglang.multimodal_gen.runtime.server_args import ServerArgs
from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger

logger = init_logger(__name__)


class WanPipeline(LoRAPipeline, ComposedPipelineBase):
    """
    Wan video diffusion pipeline with LoRA support.
    """

    pipeline_name = "WanPipeline"

    _required_config_modules = [
        "text_encoder",
        "tokenizer",
        "vae",
        "transformer",
        "scheduler",
    ]

    def initialize_pipeline(self, server_args: ServerArgs):
        # We use UniPCMScheduler from Wan2.1 official repo, not the one in diffusers.
        self.modules["scheduler"] = FlowUniPCMultistepScheduler(
            shift=server_args.pipeline_config.flow_shift
        )

    def create_pipeline_stages(self, server_args: ServerArgs) -> None:
        self.add_standard_t2i_stages()


EntryClass = WanPipeline